perf: size default ingest transport to worker count

[acoshift]

Summary

The httpClient() fallback used http.DefaultClient, whose transport caps MaxIdleConnsPerHost at 2. Each ingest worker's flush is a synchronous blocking POST, so effective throughput scales with SetConcurrent — but raising it past 2 closed every connection beyond the first two, forcing a fresh TCP (and TLS) handshake on each flush and defeating the point of more workers.

This replaces the bare fallback with a lazily-built, cached client whose transport is sized to the worker count:

• Clones http.DefaultTransport (preserves the dial/TLS/HTTP2 defaults).
• Sets MaxIdleConnsPerHost = getConcurrent() so each worker keeps its connection hot for reuse.
• Bumps MaxIdleConns only when it's below the worker count (default 100 otherwise untouched).
• Built once behind a sync.Once and shared across workers.
• No client-level Timeout — the existing per-request context deadline from getIngestTimeout() still bounds each flush, so behavior is unchanged.

A user-supplied client via SetHTTPClient still passes through verbatim.

Notes

• The pool is sized at the first httpClient() call, which in the ingest path is always after setup(), so the worker count is locked in. The only mismatch window is the unusual Search → SetConcurrent → Ingest ordering, matching the existing assumption that setters run before use.
• MaxConnsPerHost is left unlimited: each worker issues at most one in-flight request, so total connections are already bounded by the worker count without risking a worker blocking on the pool.

Test plan

• New white-box tests (client_internal_test.go): default transport sized to concurrent, default client cached across calls, user client passthrough.
• go vet ./... clean
• go test ./... pass
• go test -race ./... pass

🤖 Generated with Claude Code